Speech De-identification of Chinese, English and Minnan: Effectiveness of Chinese-based LLM Model and ASR

This study addresses the critical challenge of privacy preservation in multilingual clinical environments, specifically targeting English, Mandarin Chinese, and Minnan. We propose a robust automatic de-identification framework that integrates a multilingual pre-trained Automatic Speech Recognition model with Large Language Models to accurately transcribe and redact sensitive patient information. A key contribution of this work is the evaluation of efficient fine-tuning strategies; we demonstrate that a partial fine-tuning approach - focusing on specific model layers with limited domain-specific data—substantially enhances performance compared to full-parameter training. To overcome challenges inherent in code-switching and mixed-script scenarios, the system employs a hybrid recognition mechanism combining dictionary-based matching, regular expressions, and Large Language Models -based reasoning. Experimental evaluations validate the efficacy of this approach, where the optimized model achieved a Word Error Rate of 7.86% and a Character Error Rate of 4.17% in mixed-language tasks. These findings confirm that the proposed method provides a scalable and accurate solution for de-identifying medical speech, facilitating the secondary use of clinical data for research while ensuring strict patient privacy.